Robin Li, Chairman and CEO of Baidu has not minced words regarding the over-proliferation of AI models in China, lambasted the " war of a hundred models" as a substantial waste of societal and computational resources.
At the 2024 World Artificial Intelligence Conference (WAIC), Li pointed out that over the past year, the industry has been overly consumed by the hype for new models and updates."Shocking announcements today, epic updates tomorrow—but where are the practical applications?" Li questioned, challenging the industry's focus.
Application first
The 2024 WAIC, held in Shanghai on July 4, is regarded as the most prestigious AI forum in China. It brought together top figures from academia, industry, and research, and featured a roundtable discussion with Turing Award winners Andrew Yao, Raj Reddy, and Manuel Blum.
In his speech, Li emphasized that without practical applications, even the best foundational models, whether open-source or proprietary, are worthless. He urged the industry to focus on applications rather than models, asking, "Where are the applications? Who benefits from them?"
Li noted that applications are not far off, as foundational models are already beginning to penetrate various industries and fields. For example, the Ernie bot from Baidu sees over 200 million daily requests, with recent averages surpassing 500 million.
However, Li warned against falling into the "super app trap," where success is measured by daily active users (DAU) reaching a billion. In the AI era, he argued, "super capable" applications that generate value for industries and specific use cases are more important than having high DAU.
Rising AI start-ups
As diverse models flourish, many internet executives in China have ventured into generative AI startups. Since 2023, at least 25 executives from companies like Alibaba, Tencent, Baidu, ByteDance, JD.com, Meituan, NetEase, and Kuaishou have left to start AI businesses, forming a new wave in the AI-Generated Content (AIGC) sector.
These startups span various fields, including general large models, vertical models, generative AI applications, AI infrastructure, AI data services, and AI consulting. The 2024 WAIC showcased many of these new companies and their models, such as SenseTime's Vimi, Alibaba Cloud's Tongyi Lingjun coding assistant, Alipay's smart assistant, and Zhipu AI's A-base large model, all featured as "treasures of the conference."
Additionally, Step Fun, founded by former Microsoft VP and chief scientist of Microsoft Research Asia, Daxin Jiang, introduced the Step series models. These include the trillion-parameter language model Step-2, the hundred-billion-parameter multimodal model Step-1.5V, and the image generation model Step-1X.
China's paths to AGI
Step Fun, established in April 2023, aims for Artificial General Intelligence (AGI) by focusing on computational power, data, algorithms, and systems. The newly launched models align with the scaling law, multimodal understanding, and generation, essential for AGI's trillion-parameter and multimodal fusion goals.
The scaling law suggests a power-law relationship between model performance and the size of the model, dataset, and computational resources. As these factors increase exponentially, the model's parameter count determines its capability, making trillion-parameter models critical for AGI.
However, differing opinions have emerged. Zhilin Yang, CEO of Moonshadow, calls the scaling law a fundamental principle for AGI. Wang Xiaochuan, CEO of Baichuan Intelligence, on the other hand, believes that scaling alone is insufficient and that AGI requires paradigm shifts in large models.
Daxin Jiang of Step Fun maintains that "trillion parameters" and "multimodal fusion" are indispensable for achieving AGI.
